def test_best_response_rps():
bart_simpson_strategy = ConstStrategy((1, 0, 0))
game = RockPaperScissors()
for p in [0, 1]:
strategy = BestResponse(game, p, [bart_simpson_strategy] * 2)
assert tuple(strategy.best_responses.values())[0] == pytest.approx(
(0.0, 1.0, 0.0))
assert strategy.value == pytest.approx(1.0)
def test_rps():
g = RockPaperScissors()
us = UniformStrategy()
rng = get_rng(seed=3)
params = rng.rand(3, 3) - 0.5
vs = LinearValueStore(params, fix_mean=0.0, regularize_l1=6.0)
infosampler = InformationSetSampler(g, us)
val = SparseSGDLinearValueLearning(g, matrix_zerosum_features, vs, infosampler, seed=44)
val.compute([us, us], 100, 0.1, 0.1)
val.compute([us, us], 100, 0.01, 0.01)
val.compute([us, us], 100, 0.001, 0.001)
def test_approx_best_response_rps():
bart_simpson_strategy = ConstStrategy((1, 0, 0))
game = RockPaperScissors()
for p in [0, 1]:
s = ApproxBestResponse(game,
0, [bart_simpson_strategy] * 2,
iterations=200,
seed=23)
assert s.strategy((), 3) == pytest.approx((0.0, 1.0, 0.0))
assert s.sample_value(50) == pytest.approx(1.0)
def test_best_response_rps():
bart_simpson_strategy = FixedStrategy(Explicit([1, 0, 0], values=["R", "P", "S"]))
game = RockPaperScissors()
strategy = BestResponse(game, 0, {1: bart_simpson_strategy})
assert list(strategy.best_responses.values())[0].probability("R") == 0.0
assert list(strategy.best_responses.values())[0].probability("P") == 1.0
assert list(strategy.best_responses.values())[0].probability("S") == 0.0
assert strategy.value == pytest.approx(1.0)
strategy = BestResponse(game, 1, {0: bart_simpson_strategy})
assert list(strategy.best_responses.values())[0].probability("R") == 0.0
assert list(strategy.best_responses.values())[0].probability("P") == 1.0
assert list(strategy.best_responses.values())[0].probability("S") == 0.0
assert strategy.value == pytest.approx(1.0)
def main():
print("#### Rock-paper-scissors value estimation")
g = RockPaperScissors()
us = UniformStrategy()
infosampler = InformationSetSampler(g, us)
val = LPZeroSumValueLearning(g, infosampler, matrix_zerosum_features, us)
# Regularize: set one payoff to 1.0
val.add_condition({(0, 1): 1.0}, 1.0)
print("# With only non-triviality (one payoff set to 1.0)")
print(val.compute())
print("Flex value sum", val.flex_sum)
# Zero diagonal
for i in range(3):
val.add_condition({(i, i): 1.0}, 0.0)
print("# With zero diagonal")
print(val.compute())
print("Flex value sum", val.flex_sum)
# Symmetrical payoffs
for i in range(3):
for j in range(i):
val.add_condition({(i, j): -1.0, (j, i): -1.0}, 0.0)
print("# Adding val(i,j) = -val(j,i)")
print(val.compute())
print("Flex value sum", val.flex_sum)
#return ### Goofspiel(3) is boring, Goofspiel(4) hits OOM
print("#### Goofspiel(4) card value estimation")
g = Goofspiel(4)
mc = OutcomeMCCFR(g, seed=42)
mc.compute(2000)
ef = InfoSetExpectedFeatures(g, goofspiel_feaures_cards, mc)
for i, f in ef.info_features.items():
print("INFOSET {}:\n{}".format(i, f))
print(ef.info_next[i])
return
val = LPZeroSumValueLearning(g, infosampler, goofspiel_feaures_cards, mc)
# Regularize: set one payoff to 1.0
val.add_condition({(0, ): 1.0, (1, ): 1.0, (2, ): 1.0, (3, ): 1.0}, 10.0)
print("# Regularizing card values mean to 2.5 (mean of 1..4)")
print(len(val.conds_eq), len(val.conds_le), len(val.flex_variables))
print(
val.compute(
options=dict(tol=1e-6, disp=True, sparse=True, lstsq=True)))
print("Flex value sum", val.flex_sum)
def main():
g = MatchingPennies()
base = np.array([[1.0, 0.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0]])
plot_to_files(g, "plot_mccfr_trace_pennies", 3, 1500, 150, base=base, exploit_every=1)
g = RockPaperScissors()
plot_to_files(g, "plot_mccfr_trace_rps", 3, 1500, 150, burn=0.3, exploit_every=1)
g = Goofspiel(4, scoring=Goofspiel.Scoring.ZEROSUM)
plot_to_files(
g,
"plot_mccfr_trace_goof4",
6,
1000000,
1000,
depth=6,
burn=0.3,
burn_from=3,
exploit_every=1)
g = Goofspiel(5, scoring=Goofspiel.Scoring.ZEROSUM)
plot_to_files(
g,
"plot_mccfr_trace_goof5",
6,
1000000,
1000,
depth=6,
burn=0.3,
burn_from=3,
exploit_every=10)
g = DicePoker(6)
plot_to_files(
g,
"plot_mccfr_trace_dicepoker",
6,
500000,
500,
depth=6,
burn=0.3,
burn_from=3,
exploit_every=1)
def test_infoset():
g = RockPaperScissors()
us = UniformStrategy()
iss = InformationSetSampler(g, [us, us])
assert iss._player_dist.probs == pytest.approx(np.array([0.5, 0.5]))
assert iss._infoset_dist[0].probs == pytest.approx(np.array([1.0]))
assert iss._infoset_dist[1].probs == pytest.approx(np.array([1.0]))
assert iss._infoset_history_dist[0][()].probs == pytest.approx(
np.array([1.0]))
assert iss._infoset_history_dist[1][()].probs == pytest.approx(
np.array([1.0, 1.0, 1.0]) / 3)
iss.sample_player()
iss.sample_info()
assert iss.sample_info(0)[1] == ()
assert iss.sample_info(1)[1] == ()
assert isinstance(iss.sample_state()[2], Situation)
assert isinstance(iss.player_distribution(), Distribution)
assert isinstance(iss.info_distribution(0), Distribution)
assert isinstance(iss.state_distribution(0, ()), Distribution)